Classification of Defective Fabrics Using Capsule Networks
Round 1
Reviewer 1 Report
The topic of your study is interesting, but when I read through the article, I felt like you describe so much of other people's research that there is very little of your own. I suggest you describe more of your research.
1)In section 3.1 you have performed data augmentation on the dataset. Used augmentation techniques are flips (vertical & horizontal), different rotation angles. But you mentioned earlier that the defect of CNN is that it cannot effectively process direction information, will this lead to a decrease in the performance of CNN, so your model surpasses CNN?
2)In section 3.2.2 you mentioned that your model architecture is slightly different from the original Hinton model, I suggest the differences are better represented in the form of figures or tables.
3)You mentioned that hyperparameter settings have a very significant effect on the accurate prediction of the model, so which of your parameters have a big impact on the experimental results, and you can try to explain why.
4)The depth of the network mentioned in line 292 is helpful for improving the accuracy. As far as I know, the network of Resnet is famous for its depth, so why not use Resnet as a comparison model.
5) the experimental part is relatively simple, only comparing the CCR index results of multiple models on the TILDA dataset, and there is no ablation experiment.
In conclusion, I invite you to introduce more about your own research.
Author Response
Hello Dear Reviewer,
We wish you a healthy day. You can find our responses in the attached file.
Regards...
Author Response File: Author Response.docx
Reviewer 2 Report
The writing and presentation quality of this work is average. Abstract and Conclusion must be revised to provide a clear idea about this work and its contributions. In this study, the authors present Capsule Networks as an alternative to Convolutional Neural Networks for deep learning tasks. Compared to mainstream deep learning algorithms, this method offers improved performance in terms of accuracy. Authors have performed this method under different circumstances and reported a performance value of 98.7%. I have noticed some issue which authors must address to improve the quality of this work.
- It is suggested to rewrite this sentence “Capsule Networks are an important method that does minimize information loss”
- Authors should revise the Abstract and highlight their main contributions. They did not mention model name CapsNets, TILDA images, model training over 270 epochs etc.
- In the introduction section, the authors should add more significant discussion about CapsNets. The reference literature must be provided from the latest research contributions.
- Table 1, authors should carefully check each abbreviation and define it at the first place of appearance.
- In TILDA dataset, authors have considered 692 or 682 images? As authors consider the first five classes (nondefective, hole, stain, thread, weave) have 100 pictures in each class, and the knitting and shaded classes have 94 and 88 pictures, respectively.
- Authors should carefully check that equations and variables are written properly as it seems there are some typing problems and variables are not fully clear.
- Table 2, the authors reported the hyperparameters used in their model without providing any details about choosing these specific parameters.
- Insufficient discussion is provided in section 3.3. Authors should provide a more comprehensive discussion about experiments and findings for better understanding and readability.
- In References, most of the studies are old. Thus, there is a clear lack of reference literature from very recent years. It’s suggested to add more references from very recent research contributions conducted in 2020-2021.
In my opinion, this study lacks reference literature, experiments, and results discussion. It does not justify considerable research contribution as authors have taken some datasets and used the previous model to carry out some analysis. The authors could not completely describe the selection of each parameter with sufficient detail.
Author Response
Hello Dear Reviewer,
We wish you a healthy day. You can find our responses in the attached file.
Regards...
Author Response File: Author Response.docx
Reviewer 3 Report
This paper presents an interesting idea to leverage textile with machine learning. There are some findings from this paper.
- It is true that CNN loses information for pooling. But there are different models that do not have pooling operations after many layers. The data loss can be mitigated by proper training. How many layers do you have in the proposed capsnet? Also how many convolutional layers were in the CNN?
- Can you provide some performance comparison with some prior works?
- You mentioned the 2nd problem is CNN needs a larger dataset. Why larger data set is a problem in this specific example? (Line 211)
- In the following line, you also mentioned the third problem is poor imagination of human visual systems. I do not agree that it is a problem. Human has a great visual system and we try to leverage that visual system in the circuit. The problem is we do not apply our attention to that level. Please cite this paper to explain the beauty of human visual systems and how attention based on spatiotemporal saliency can do more to improve the overall system. https://link.springer.com/article/10.1007/s11554-020-00960-5
Author Response
Hello Dear Reviewer,
We wish you a healthy day. You can find our responses in the attached file.
Regards...
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Thanks to the authors for revising the article accordingly. All of my concerns have been resolved.
Reviewer 2 Report
Thank you for your explanation of each question. You have addressed all my concerns very well in this revised manuscript. The quality of this work is improved.